Agentic TurnsIt sounds sure when it's wrong
AI models can invent a price, a policy or an opening hour and say it with total confidence. One wrong answer to the wrong customer costs more than the tool saves.
AI agent and model testing · Sydney
Test your AI before your customers do.
If a chatbot, AI receptionist, voice agent or AI feature talks to your customers, it can make things up, leak details or say yes to the wrong thing. Agentic Turns tests it against your real situations, one conversation turn at a time, and tells you, in plain English, what's safe to switch on and what needs fixing first.
- Works with any model or vendor
- Tested on your own cases, not generic ones
- Every finding backed by evidence
A demo that works isn't proof it's ready.
People will try to trick it
Customers, competitors and bots will try to talk your AI into discounts, rude replies or revealing what it shouldn't. Most small-business AI has never been tested for this.
Updates change behaviour
Vendors update their models and your team tweaks the instructions. Something that worked last month can quietly break. Without a test set, nobody notices until a customer does.
What we test
Ten areas, chosen to match the risks that matter for your business. A Health Check covers the five that matter most for you; a Launch Test covers all ten.
Accuracy
Does it answer your real customer questions correctly and usefully? Scored against answers you agree are right.
Made-up answers
Does it invent prices, policies, bookings or facts? Does it stick to your own information and say "I don't know" when it should?
Safety and misuse
Red-team testing: tricks, jailbreaks, hidden instructions in emails or web pages, and attempts to make it rude, unsafe or off-brand.
Privacy and data leaks
Can someone get it to reveal another customer's details, its hidden instructions, or keys? Checked against the Australian Privacy Principles.
Fairness and bias
Does it treat people the same whatever their name, age, accent, suburb or way of writing? We swap only those details and compare the answers.
Agent control
For AI that takes actions: does it stay inside its permissions, ask before anything it can't undo, stop when told and avoid runaway loops?
Voice and call rules
For voice agents: does it say it's an AI, mention recording, end the call when asked, keep to calling hours and respect do-not-call requests?
Cost, speed and uptime
What each conversation really costs, how long people wait, and how often it times out or fails. Measured, not estimated.
Change and regression
We save your test set and rerun it after any model update or prompt change, so you can see exactly what got better or worse.
Model choice
Which model gives you the results you need for the lowest cost? We compare the options side by side on your own cases.
What we can test
If it uses an AI model and talks to people or makes decisions, we can test it. It doesn't matter who built it or which model it runs on.
- Customer-facing: website chatbots, AI receptionists and voice agents, SMS and email assistants, booking bots.
- Inside the business: AI that drafts quotes, emails or reports, sorts enquiries, or answers staff questions from your documents.
- For builders: agencies and developers who want independent evidence before they hand an AI product to their client.
How a test works
- 1
Scope
A free 20-minute call. What the AI does, who it talks to, what could go wrong, and what "good" looks like for you.
- 2
Build your test set
We write test cases from your real situations, plus tricky and hostile ones. You approve the expected answers. Some cases are held back so nothing is tuned to pass.
- 3
Run and review
Automated runs score every case, and a person reviews the failures and the borderline ones. A number alone is never the verdict.
- 4
Report
A pass, watch or fail for each area, the problems ranked by how much they matter, and the exact fix we'd make for each.
- 5
Fix and retest
Once you or your developer make the fixes, we rerun the same test set and show what changed.
What you get
- A one-page scorecard you can show your team, your client or your insurer.
- An evidence log with every test case, what the AI said and why it passed or failed.
- A ranked fix list with what to change in the instructions, data, settings or model.
- Your test set to keep, so the same checks can be rerun after every change.
| Area | Result |
|---|---|
| Accuracy | Pass |
| Made-up answers | Watch |
| Safety and misuse | Fail |
| Privacy and data leaks | Pass |
| Cost per conversation | Pass |
Each result links to the cases behind it.
Packages
Fixed scope and a fixed price, agreed in writing after the free scoping call. No hourly surprises.
Health Check
One-off
One AI tool, the five areas that matter most for you, about 60 test cases. Report within 5 business days, with one retest.
Best for: an AI tool you already use and want to trust.
Launch Test
One-off
All ten areas, including red-team, privacy and agent-control testing, up to 250 test cases, a fix session with your team and a full retest.
Best for: a new chatbot, voice agent or AI product.
Model Bake-off
One-off
Up to four models compared on 100 of your own cases for accuracy, cost and speed, with a clear recommendation.
Best for: choosing a model, or cutting the AI bill.
Ongoing Assurance
Monthly
Your test set rerun every month and after any model or prompt change, an alert when results drop, and a short monthly report.
Best for: AI that's live and changing.
Testing for health, disability, finance or legal settings is scoped separately, because the stakes and the rules are higher.
Built on recognised guidance
Our test areas are mapped to published guidance, so your report lines up with what boards, clients and insurers are starting to ask for.
Agentic Turns isn't a certification body. Our report is independent evidence for your own decisions, not a certificate or a legal opinion.
- OWASP Top 10 for LLM Applications and the OWASP Top 10 for Agentic Applications: the common ways AI apps and agents get attacked or go wrong.
- NIST AI Risk Management Framework: how to measure and manage AI risk.
- Australia's Guidance for AI Adoption and the Australian Privacy Principles: the local expectations for responsible AI and personal information.
We hold our own AI to the same standard.
Agentic Turns is run by Ashu Nand in Sydney, who also runs AI agents in his own service businesses, including Bright Nest Cleaning. Those agents work under written limits, approval steps for anything that reaches a customer, and an off switch. The testing method we sell is the one used on those agents before they change.
"No agent marks its own work. A change goes live only after it's tested on cases it hasn't seen, reviewed by someone else, and approved."
From the operating rules we test our own agents by.Questions
Do you need access to our systems?
Usually not. Most tests run through the same chat window, phone number or test link your customers would use. If we need more, we ask for a test account with the least access that does the job, and we hand it back when we finish.
What happens to our data?
We ask for made-up or de-identified examples wherever possible. Anything you give us is used only to test your system, kept only as long as the job needs, and deleted when you ask. See our privacy policy.
Will you fix the problems too?
The report tells you exactly what to change. Your own developer or supplier makes the fixes, and the retest is included. We don't do the fixing ourselves, so the verdict stays independent.
Which AI models do you test?
Any of them: the big commercial models, open models, or a product built on top of one. We test what your customers actually experience, whoever built it.
Can you guarantee our AI won't make mistakes?
No, and nobody honestly can. Testing finds problems before customers do and shows how often things go wrong. Ongoing Assurance keeps checking as things change.
Why is it called Agentic Turns?
A turn is one back-and-forth in a conversation. AI agents rarely fail on the first question. They slip a few turns in, when a customer pushes back, changes their mind or tries a trick. So we test whole conversations, turn by turn, not just single questions.
Do you build AI as well?
No. Agentic Turns only tests, so we have no reason to go easy on anyone's work. If you want an AI receptionist built, our sister business NAND AI does that, and we test its work the same way as anyone else's.
Book a free 20-minute scoping call
Tell us what your AI does and what worries you about it. Ashu will be in touch within one business day with a time and the questions we'll cover.
Loading the form…